Does AI control nuclear weapons? Who keeps the button

Who has launch authority today?

Nuclear states say little about their command systems, so the public record is thin. Here is what each one says in its own documents, where we could find it.

• -United States. "The U.S. President has sole authority to authorize the use of U.S. nuclear weapons," according to the Congressional Research Service. Military leaders advise, and are then "required to transmit and implement the orders" if the President decides to use nuclear weapons.1 CRS also notes the time pressure the system was built for: missiles from Russia could reach the US in around 30 minutes, and some analysts estimate the President would have less than 10 minutes to decide in that scenario.1
• -United Kingdom. "Only the Prime Minister can authorise the use of our nuclear weapons," says the government's current guide to the deterrent, which also says the UK deterrent "remains fully operationally independent" although it is declared to NATO.2
• -France. A Ministry of the Armed Forces guide says the President is assured that nuclear forces can be used only according to his directives and that "lui seul peut leur donner l'ordre ultime d'engagement" ("he alone can give them the final order to engage", our translation).3
• -Russia. Russia's 2024 nuclear doctrine, signed by President Putin on 19 November 2024, says (in a US Naval War College informal translation) that "The President of the Russian Federation makes the decision on the use of nuclear weapons."4
• -India. A government release of 4 January 2003 says India's Nuclear Command Authority has a Political Council chaired by the Prime Minister, and that "It is the sole body which can authorize the use of nuclear weapons."5
• -China, Pakistan, Israel, North Korea. We did not find an official, current public statement of who authorises use that we could open and cite. Pakistan's arrangements have changed several times, and older accounts may be out of date. This page does not guess.

What have governments said about keeping humans in control?

The pledges below are about the decision to use nuclear weapons. The Arms Control Association (ACA) collects them and calls the result "a broad but loose consensus for ensuring a 'human in the loop'".6

• -US, UK and France (2022). In a joint paper to the 2022 Non-Proliferation Treaty conference: "we will maintain human control and involvement for all actions critical to informing and executing sovereign decisions concerning nuclear weapons employment".8
• -US Nuclear Posture Review (October 2022). "In all cases, the United States will maintain a human 'in the loop' for all actions critical to informing and executing decisions by the President to initiate and terminate nuclear weapon employment."9 The current administration repeated the commitment in its April 2026 report to the NPT conference.10
• -UK (June 2022). The Defence AI Strategy: "regardless of any use of AI in our strategic systems – human political control of our nuclear weapons is maintained at all times."11
• -US and China (November 2024). After their meeting in Lima, the White House readout said Presidents Biden and Xi "affirmed the need to maintain human control over the decision to use nuclear weapons."7
• -China (November 2025). A government white paper says, about military AI in general, that "all relevant weapons systems are under human control to prevent unauthorized actions."12
• -Russia (2026). Its report to the NPT conference says it "ensures decisive human involvement and judgment in the field of nuclear weapons on an ongoing basis" (UN English translation).23
• -India and Pakistan. India's Foreign Secretary told the Conference on Disarmament in February 2026 that decisions on nuclear weapons will remain under human control (as reported by All India Radio, which paraphrases him).24 Pakistan says it supports the principle of human control over nuclear-use decisions and "urges all nuclear-weapon states to endorse" it.25
• -US Congress (December 2024). The defence authorisation act says it is US policy that AI "should not compromise the integrity of nuclear safeguards", including "the principle of requiring positive human actions in execution of decisions by the President with respect to the employment of nuclear weapons."13

What the pledges leave open: ACA notes there are "no agreed definitions, constraints, or guidelines on how AI tools might be utilized in nuclear weapons operations", and no active talks among the major nuclear powers on how to put the principle into practice.6

Where is AI used around nuclear command, and where is it not?

"Nuclear command and control" covers far more than the final order: early-warning satellites and radars, intelligence, communications and planning. Some of it has been automated for decades.1518

Where AI is used or planned (as officials and researchers describe it):

• -Sorting data and warnings. A 2020 SIPRI study said machine learning and autonomy "hold major promise for early warning" and intelligence gathering: spotting objects and patterns in large data sets and helping commanders see sooner whether an attack may be under way. It concluded that this "could give a human military commander better situational awareness and potentially more time to make decisions."15
• -Decision support. In October 2024 US Strategic Command's then commander, Gen. Anthony Cotton, said "AI will enhance our decision-making capabilities" and "Advanced systems can inform us faster", but that "we must always maintain a human decision in the loop".14 He also called for research into the risks of "indirect integration of AI into nuclear decision-making processes".14
• -Networks and planning. The US Congress says AI may help "strengthen the security of critical strategic communications and early warning networks" and improve planning, as long as it does not "increase the risk" that strategic assets are compromised.13

Where it is not, by stated policy: the decision itself. Every statement above keeps a person in charge of launching. SIPRI also wrote that there "seems to be a general agreement among nuclear-weapon experts that machine learning and autonomy should not be integrated into nuclear command and control", partly because today's algorithms are too unpredictable for a system that is "too safety critical".15 Public detail on how much automation each state uses is sparse, so outsiders can't fully check this.18

What risks do experts name? Speed, false alarms and over-trust

Researchers who study this list a few recurring risks.1518

• -Speed. If AI makes warfare faster, decision time shrinks. SIPRI warns that "the more that human decision time is compressed and automation is integrated into command and control, the greater becomes the risk of an accidental escalation".15
• -Over-trust ("automation bias"). People can "over-rely on a system and assume that the information provided by the system is correct". The opposite, under-trust, is also a risk.15
• -Bad or poisoned data. An opponent could try "data poisoning to undermine or manipulate the performance of early-warning systems".15
• -A human who is in the loop only on paper. The Bulletin of the Atomic Scientists notes the open question of "how to avoid situations where humans formally retain the final decision but, in practice, are heavily dependent on opaque AI systems."18

Why false alarms matter: the history. Warning systems have been wrong before, and people caught it:

• -9 November 1979 (US). A war-game test tape was mistakenly put into a NORAD computer, showing a Soviet attack that was not happening. Officers checked the raw satellite and radar data, saw nothing, and cancelled the alert.1516
• -3 June 1980 (US). Warning displays again showed an attack; the cause was traced to a failed computer chip that the National Security Archive's documents put at 46 cents.16
• -26 September 1983 (USSR). The Soviet early-warning system reported five US missiles. Lt. Col. Stanislav Petrov judged it a false alarm and reported it as one; as SIPRI recounts it, he reasoned that "when people start a war, they don't start it with only five missiles." The satellite had misidentified the missiles.15

In each case a person slowed down and checked. That is the argument for keeping humans in the loop, and the reason experts worry about anything that pushes them to decide faster.115

What about AI war games?

In February 2026 King's College London announced a study by Professor Kenneth Payne that put three AI language models (GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash) into 21 simulated nuclear crises. All 21 games featured nuclear signalling (threats) by at least one side, and "none of the models ever chose accommodation or surrender."22 This was a simulation of chatbots playing leaders, not a test of any real weapons system, and no government has said it uses such models in nuclear decisions. Payne called the results "sobering".22

Are there pending bills or UN talks?

• -US bills. The Block Nuclear Launch by Autonomous Artificial Intelligence Act (H.R. 2894, introduced 26 April 2023 by Rep. Ted Lieu with Reps. Ken Buck and Don Beyer; Senate version S. 1394) would bar federal money for launching a nuclear weapon with "an autonomous weapons system that is not subject to meaningful human control".20 It was not passed in that Congress, and we found no reintroduction under that name in the current (119th) Congress as of October 2026. The policy statement in the defence law signed in December 2024 covers similar ground.13 Separately, CRS lists 2025 bills on presidential launch authority itself, such as the Restricting First Use of Nuclear Weapons Act of 2025 (H.R. 669 / S. 192).1
• -UN General Assembly. In December 2024 the Assembly asked the Secretary-General to gather states' views on AI in the military domain (resolution 79/239).19 On 1 December 2025 it adopted its first resolution on AI in nuclear command systems,18 demanding that "human control and oversight is maintained over command, control and communications systems of nuclear weapons, including those that integrate artificial intelligence technology". The vote was 118 for, 9 against (including France, Russia, the UK and the US) and 44 abstentions.17 Russia explained its vote by the lack of agreed definitions of "artificial intelligence" and "meaningful human control" (as reported by the Bulletin).18
• -NPT Review Conference. The April–May 2026 conference ended on 22 May without an agreed final document, the third in a row; disputes over Iran were a main reason.21 So there is no new treaty-level text on AI and nuclear weapons.
• -What to watch. The UN informal exchanges in Geneva on military AI decided in December 2025,17 and the next steps on autonomous weapons under the UN weapons convention (CCW), covered in Are killer robots real?.

FAQ

Will AI nuclear weapons kill us?

No AI controls a nuclear launch under any published policy. The realistic risks experts discuss are mistakes in warning and decision support, which is why the pledges focus on human control.615

Could AI hack into nuclear weapons?

We found no public report that this has happened. Congress and SIPRI both flag cyber and data risks as reasons for care.1315

Does the "Dead Hand" mean Russia's nukes are automatic?

Not covered here: we found no official document to cite.